Post on 08-Jan-2018
description
Estimating the uncertainties in the products of inversion algorithms or, how do we set the error bars for our inversion
results?Emmanuel Boss, U. of Maine.
•In science there are no quantities that have no uncertainties associated with them.
•The way we present uncertainties graphically is with error bars.
•Sometimes error bars are either too small to be noticed or simply neglected.
•In the least, uncertainties should be reported so significance in reported relationship can be evaluated.
Sources of uncertainties in empirical inversion algorithms:
•Uncertainties in the training set (e.g. biases in the data set).
•Data inverted not covered by the training set (application of open ocean algorithm in coastal environment).
•Uncertainties in the inverted data (e.g. uncertainties in value of input reflectance).
•Differences is sensor-specific algorithms derived from common training sets (SeaWiFS has different wavelengths than MODIS, thus OC4 and OC3M are slightly different)
Sources of uncertainties in semi-analytical algorithms:
•Uncertainties in the relationship between Rrs and IOPs (e.g. BRDF, non-elastic scattering).
•Uncertainties in assumed shapes of IOPs (e.g. phytoplankton absorption).
•Uncertainties in the inverted data (e.g. uncertainties in the value of Rrs()).
•Uncertainties in the cost functions (weights) applied to the statistical inversion method.
Quantifying uncertainties in empirical inversion algorithms:
•Use a testing data set collected in the environment of interest to evaluate the likely uncertainties of the inversion algorithm (e.g. how well can we obtain [chl] for the Gulf of Maine in January from SeaWIFS?).
•Quantify the statistics*of the difference between inverted value and measured value to obtain:
•Bias- how accurate are the inverted values on average? – if a bias exist, re-evaluate the inversion parameters.
•Precision- what is the absolute (or relative) difference between predicted values and inverted values? – use this as your estimate for error bars.
* If the underlying statistics are not known nonparametric statistics are safest.
What do you do if you don’t have a testing set?
Quantifying uncertainties in semi-analytical algorithms:
Part I: Uncertainties in the relationship between Rrs and IOPs.
Use a testing set (IOP and Rrs measured close by) to validate the inversion where Cloud cover/sun angle/surface conditions varied in an area with relatively constant IOP.
Evaluate the contribution of inelastic scattering by incorporating the following iteration (inspired by Pozdnyakov and Grassl, 2003):
1. Once you obtained aCDOMand aphyto, recalculate their likely fluorescence (as well as the contribution of Raman scattering), by e.g., inputing the IOP retrieved into Hydrolight. Note: one needs to choose the fluorescence quantum yields for CDOM and CHL. Residuals will help with estimating chl.
2. Rrs_new=Rrs_inverted-(Rrs_Hydrolight-Rrs_inverted). Invert again, until a convergence criteria is reached (e.g. values change by less than x% between iterations).
Quantifying uncertainties in semi-analytical algorithms:
Part II: Uncertainties in assumed shapes of IOPs (e.g. phytoplankton absorption, spectral slopes of backscattering and aCDM).
A sensitivity analysis is performed looking at how the output varies with shapes of IOPs (e.g. Roesler and Perry, 1995, Lee et al., 1996, Hoge and Lyons, 1996, Garver and Siegel, 1997).
Roesler iterative method for optimizing a(): look at the residuals (Rrs_measured-Rrs_modeled). If they look like pigments peaks, modify a to include pigment in those wavelengths (e.g. from a library of spectra one has established ahead of time).
Quantifying uncertainties in semi-analytical algorithms:
Part III: Uncertainties in the inverted data (e.g. uncertainties in the value of Rrs()).
22
d
d
u
u
rs
rs
d
urs E
ELL
RR
ELR
Note: For Rrs obtained from a satellite, the uncertainties are very likely to vary spectrally (e.g. due to atmospheric correction) and with time (e.g., due to spectrally dependent sensor drift over its mission life).
Quantifying uncertainties in semi-analytical algorithms:
The Wang/Boss/Roesler approach:
Inversion Scheme
Output:Inverted IOPs (apg(λ), bbp(λ),
aph(λ), adg(λ)) & Uncertainties in inverted IOPs
Estimate Uncertainty of
criterion of Inversion
Uncertainties in Shapes of
IOPs
Uncertainties in IOP- rrs(λ)
Relationship
Input rrs(λ) Uncertainties in rrs(λ) Measurement
(nonexistence in simulated dataset)
Quantifying uncertainties in semi-analytical algorithms:
The Wang/Boss/Roesler approach:
400 450 500 550 600 6500
0.5
1
1.5
2
2.5
3
3.5
basi
s ve
ctor
aph
400 450 500 550 600 6500
0.5
1
1.5
2
2.5
basi
s ve
ctor
adg
400 450 500 550 600 6500.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
basi
s ve
ctor
bbp
Shape of component IOPs:
Phytoplankton: Ciotti et al., 2002.acm with exponential slope varying from 0.01 to 0.02bbp with spectral slope varying from 0 to 2.
Choose one combination of shapes and invert linearly (total of 113=1331 combinations). Solve for 3 amplitudes (bbp, acm, a), for each choice of 3 shape parameters.
Quantifying uncertainties in semi-analytical algorithms:
The Wang/Boss/Roesler approach:Evaluating uncertainties
400 450 500 550 600 6500
5
10
15
20
25
wavelength [400:10:650]
rela
tive
diffe
renc
e at
eac
h w
avel
engt
h (%
)
relative difference for aphrelative difference for bbp relative difference for adg
relative difference for aphrelative difference for bbprelative difference for adg
1. Uncertainties in shapes of IOPs (ZP Lee dataset)
400 450 500 550 600 6500
2
4
6
8
wavelength [412:1:650]
rela
tive
diffe
renc
e at
eac
h w
avel
engt
h (%
)
2. Uncertainties in relation of IOPs and rrs.
3. Stability of radiometers<2%
Any solution within 10% rrs at all is kept.
400 450 500 550 600 6502
4
6
8
10
wavelength [400:10:650]
rela
tive
diffe
renc
e at
eac
h w
avel
engt
h (%
)
{Hyd
rolig
ht-f(
IOP
)}{0
.5(H
ydro
light
+f(IO
P)}
Fiel
d: %
Diff
(rrs)
whe
n IO
P=c
onst
.
Quantifying uncertainties in semi-analytical algorithms:
The Wang/Boss/Roesler approach, example:
450 500 550 600 6500
1
2
3
4
5
6
7
8
x 10-3
wavelength [412:1:650]
inpu
t rrs
& m
edia
n in
vert
ed rr
s w
ithin
10%
0.04 0.045 0.05 0.0550
10
20
30
40
50
distribution of apg(440)1.7 1.8 1.9 2 2.1 2.2 2.3
x 10-3
0
10
20
30
40
50
60
70
distribution of bbp(555)
0 0.01 0.02 0.03 0.040
10
20
30
40
distribution of aph(440)0.01 0.02 0.03 0.04 0.05 0.060
10
20
30
40
50
distribution of adg(440)
Quantifying uncertainties in semi-analytical algorithms:
The Wang/Boss/Roesler approach: Results for 31 field matchups
10-1 100
10-1
100
measured apg at 440
inve
rted
apg
at 4
40
measurement uncertainty
90% error bound
Triangle: measured IOPs by HyCODE2000 slowdrop Circle: measured IOPs by HyCODE2000 suitcase
Square: measured IOPs by HyCODE2001 slowdrop
10-3 10-210-3
10-2
measured bbp at 555
inve
rted
b bp a
t 555
10-2 10-1 10010-2
10-1
100
measured ag at 440
inve
rted
adg
at 4
40
10-4 10-3 10-2 10-110-4
10-3
10-2
10-1
measured aph at 676
inve
rted
aph
at 6
76
apg(440) bbp(555)
adg(440) a(676)
Note the error bars!!!
Quantifying uncertainties in semi-analytical algorithms:
The Wang/Boss/Roesler approach: Results for 31 field matchups
0.01 0.015 0.02 0.0250.01
0.015
0.02
0.025
measured Sg
inve
rted
Sdg
Triangle: measurement IOPs at 440nm by HyCODE2000 slowdrop Square: measurement IOPs at 440nm by HyCODE2001 slowdrop
: 1:1 line
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
measured Cp slope
inve
rted
S f
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
measured Y
inve
rted
Y
Measured sg
Inve
rted
s dg
Cio
tti’s
siz
e pa
ram
ter
Spectral slope of cp ()
Measured spectral slope of bbp
Inve
rted
Y
Roesler and Boss, 2008
Other possible eigenfunctions:
Quantifying uncertainties in semi-analytical algorithms:
Other approaches have also been explored:
Goodness of fit (& covariance matrices) in modeled vs. measured Rrs (Maritorena et al. 2010)
Propagation of error (Lee et al. 2010) (but not all errors are included).
Stratification of uncertainties by optical water type (Moore et al. 2010)
Quantifying uncertainties, summary:
• It is possible to put error bars on inversion products, so lets do it (and on global scales provide maps of uncertainties).
• Error bars will get smaller the more we know about the environment (e.g. limit the shape of the component IOP).
• The magnitude of the uncertainty may make the data useless for some application while still very useful for other.
• Another approach is to use the difference between observed rrs and that from the inverted parameters to derive the uncertainties. This approach fails to take into account the inherent uncertainties in the: 1. measured rrs, 2. IOP-rrs relationships and 3. assumed shapes for IOPs.
• A fundamental difference between the empirical and semi-analytical approaches is that one is a statistical interpolation scheme, while the other is based on the fundamental physics of remote sensing supplemented by empirical knowledge of component IOP shape.
Some useful Links and references:
Press W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. 1988. Numerical Recipes, Cambridge University Press.
Taylor, J., 1996. An Introduction to Error Analysis: The Study of Uncertainties in Physical Measurements. University Science Books.
http://en.wikipedia.org/wiki/Propagation_of_uncertainty
http://badger.physics.wisc.edu/lab/manual/node4.html